Checkpoint blockade immunotherapy exhibits durable responses in a subset of advanced stage non-small cell lung cancer (NSCLC) patients. However, not all patients respond and markers that predict response to immunotherapy are still an unmet clinical need. In this study, pre-treatment quantitative image-based biomarkers (radiomics) were utilized to predict survival risk among NSCLC patients treated with immunotherapy. Study cohorts were compromised of NSCLC patients that were treated with single or double agent immunotherapy and were divided into three independent groups from two different hospitals for Training (N = 180), test (N = 90) and validation (N = 62). Overall survival (OS) and progression-free survival (PFS) were used as endpoints. The most prognostic radiomic features and clinical covariates were used as inputs to a Classification and Regression Tree (CART) machine learning algorithm to stratify patients into survival risk-groups utilizing the training cohort. The risk groups identified were then tested and validated in the test and validation cohorts. Additionally, four independent non-immunotherapy treated NSCLC cohorts (total N = 446) were utilized to further investigate the prognostic value of the derived radiomics signature. The biological underpinnings of the most informative radiomics were assessed using gene expression data from a radiogenomics dataset and an immunohistochemistry (IHC) data.The final parsimonious model was able to stratify patients into four risk groups (from low-risk to very-high-risk) and the models were successfully tested and validated in independent cohorts. Specifically, the very-high-risk group was found to be associated with extremely poor OS outcomes (0% 3-year OS; hazard ratio [HR] = 5.35, 95% confidence interval [CI]: 2.14 - 13.36) compared to the low-risk group (38.9% 3-year OS; HR = 1.00). When PFS was assessed, similar outcomes were observed (0% vs. 29.8% 3-year PFS). A radiomic feature that quantifies the heterogeneity of tumor structural phenotype (GLCM inverse difference) was highly associated with OS in all three immunotherapy treated cohorts. Utilizing gene expression data, GLCM inverse difference was found to be positively associated with expression of carbonic anhydrase-9 (CA-IX), a surrogate for tumor hypoxia. This was validated by IHC, and further analysis showed that it was also associated with OS in the four other independent non-immunotherapy treated NSCLC cohorts.In this study, we utilized non-invasive image-based features and standard-of-care clinical covariates to identify novel risk groups that can predict OS and PFS among NSCLC patients treated with IO. These models have important translational associations as a potential risk score that could be used to assist decision making by identifying potentially vulnerable patients in the setting of immunotherapy.

Citation Format: Ilke Tunali, Yan Tan, Jhanelle E. Gray, Evangelia Katsoulakis, Steven A. Eschrich, James Saller, Theresa Boyle, Jin Qi, Albert Guvenis, Robert J. Gillies, Matthew B. Schabath. Hypoxia-related radiomics predict checkpoint blockade immunotherapy response of non-small cell lung cancer patients [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 5806.